Research map / Point clouds / Segmentation and detection
Voxel Mamba: Group-Free State Space Models for Point Cloud based 3D Object Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Voxel Mamba, a group-free state-space-model backbone that serializes all voxels into one Hilbert-ordered sequence, with dual-scale SSM blocks and implicit window positional embeddings for 3D LiDAR detection.
Problem
Serialization-based Transformers for voxel 3D object detection group voxels into short windows, which loses spatial proximity between neighboring voxels because Transformer cost is quadratic in group size.
Approach
Use SSMs' linear complexity to treat all voxels as a single group ordered by a Hilbert curve, add a dual-scale SSM block (high-resolution forward branch, downsampled backward branch) to enlarge the effective receptive field, and encode window positions implicitly via positional embeddings instead of explicit partitioning.
Key results
- Waymo validation: 79.6 L1 3D mAPH and 73.6 L2 mAPH, vs. DSVT-Voxel 78.2 / 72.1.
- Waymo test: 79.6 L1 / 74.3 L2 mAPH, better than the 3-frame PillarNeXt (79.0 / 74.1) and SST.
- nuScenes validation: 71.9 NDS and 67.5 mAP; nuScenes test: 73.0 NDS and 69.0 mAP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use group-free Voxel Mamba instead of DSVT-Voxel | Waymo Open Dataset validation (single-frame) · Waymo Open Dataset validation (single-frame) | ALL 3D mAPH L1/L2: 79.4/73.4 -> 79.6/73.6; +1.4 and +1.5 better than DSVT-Voxel | Helped |
| Use Voxel Mamba instead of the previous best method | nuScenes validation · nuScenes validation | NDS and mAP: +0.5 NDS; +0.8 mAP | Helped |
| Use Voxel Mamba instead of DSVT on nuScenes test | nuScenes test · nuScenes test | NDS and mAP: 72.7 -> 73.0 NDS; 68.4 -> 69.0 mAP | Helped |
| Use Hilbert Curve instead of random serialization (ablation) | Voxel Mamba · nuScenes validation | mAP: 66.0 -> 67.5 | Helped |
| Use Hilbert Curve instead of Window Partition (ablation) | Voxel Mamba · nuScenes validation | mAP: 67.3 -> 67.5 | Helped |
| Use Hilbert Curve instead of Z-Order Curve (ablation) | Voxel Mamba · nuScenes validation | mAP: 67.0 -> 67.5 | Helped |
| Add bidirectional SSMs and a Hilbert-based group-free sequence (ablation) | Centerpoint-Pillar baseline · nuScenes validation | mAP: 63.3 -> 65.8 | Helped |
| Convert pillar representation to voxels (ablation) | Centerpoint-Pillar with bidirectional SSMs · nuScenes validation | mAP: 65.8 -> 66.3 | Helped |
| Add Dual-scale SSM Block (DSB) (ablation) | Voxel detector with bidirectional SSMs · nuScenes validation | mAP: 66.3 -> 66.7 | Helped |
| Add Implicit Window Partition (IWP) (ablation) | Voxel Mamba with DSB · nuScenes validation | mAP: 66.7 -> 67.5 | Helped |
| Set DSB downsampling rates to {1,2,2} instead of {1,1,1} (ablation) | Voxel Mamba · nuScenes validation | mAP: 66.6 -> 66.9 | Helped |
| Set DSB downsampling rates to {1,2,4} instead of {1,2,2} (ablation) | Voxel Mamba · nuScenes validation | mAP: 66.9 -> 67.5 | Helped |
Limitations
- Curve templates in the Hilbert Input Layer consume GPU memory, which grows with voxel resolution.
- Downsampling and upsampling operations are not elaborately optimized, which limits efficiency.
Experimental setup
Datasets: Waymo Open Dataset nuScenes
Benchmarks: Waymo Open Dataset nuScenes
Metrics: mAP mAPH NDS FPS GPU memory
Baselines: DSVT-Voxel PV-RCNN++ CenterPoint-Pillar PillarNeXt SST VoxSeT PTv3 VoxelNeXt HEDNet FSDv2
Builds on: Mamba Vision Mamba DSVT OpenPCDet CenterPoint
Closest papers by meaning
- 3DET-Mamba: Causal Sequence Modelling for End-to-End 3D Object Detection (NeurIPS 2024 · similarity 0.89)
- UniMamba: Unified Spatial-Channel Representation Learning with Group-Efficient Mamba for LiDAR-based 3D Object Detection (CVPR 2025 · similarity 0.89)
- Voxel or Pillar: Exploring Efficient Point Cloud Representation for 3D Object Detection (AAAI 2024 · similarity 0.87)
- PointMamba: A Simple State Space Model for Point Cloud Analysis (NeurIPS 2024 · similarity 0.87)
- LION: Linear Group RNN for 3D Object Detection in Point Clouds (NeurIPS 2024 · similarity 0.87)
- Spatial-Mamba: Effective Visual State Space Models via Structure-Aware State Fusion (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.